Bayesian Optimization with Expected Improvement: No Regret and the Choice of Incumbent

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Jingyi, Wang, Haowei, Ng, Szu Hui, Petra, Cosmin G.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909747708428288
author Wang, Jingyi
Wang, Haowei
Ng, Szu Hui
Petra, Cosmin G.
author_facet Wang, Jingyi
Wang, Haowei
Ng, Szu Hui
Petra, Cosmin G.
contents Expected improvement (EI) is one of the most widely used acquisition functions in Bayesian optimization (BO). Despite its proven empirical success in applications, the cumulative regret upper bound of EI remains an open question. In this paper, we analyze the classic noisy Gaussian process expected improvement (GP-EI) algorithm. We consider the Bayesian setting, where the objective is a sample from a GP. Three commonly used incumbents, namely the best posterior mean incumbent (BPMI), the best sampled posterior mean incumbent (BSPMI), and the best observation incumbent (BOI) are considered as the choices of the current best value in GP-EI. We present for the first time the cumulative regret upper bounds of GP-EI with BPMI and BSPMI. Importantly, we show that in both cases, GP-EI is a no-regret algorithm for both squared exponential (SE) and Matérn kernels. Further, we present for the first time that GP-EI with BOI either achieves a sublinear cumulative regret upper bound or has a fast converging noisy simple regret bound for SE and Matérn kernels. Our results provide theoretical guidance to the choice of incumbent when practitioners apply GP-EI in the noisy setting. Numerical experiments are conducted to validate our findings.
format Preprint
id arxiv_https___arxiv_org_abs_2508_15674
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Bayesian Optimization with Expected Improvement: No Regret and the Choice of Incumbent
Wang, Jingyi
Wang, Haowei
Ng, Szu Hui
Petra, Cosmin G.
Machine Learning
Expected improvement (EI) is one of the most widely used acquisition functions in Bayesian optimization (BO). Despite its proven empirical success in applications, the cumulative regret upper bound of EI remains an open question. In this paper, we analyze the classic noisy Gaussian process expected improvement (GP-EI) algorithm. We consider the Bayesian setting, where the objective is a sample from a GP. Three commonly used incumbents, namely the best posterior mean incumbent (BPMI), the best sampled posterior mean incumbent (BSPMI), and the best observation incumbent (BOI) are considered as the choices of the current best value in GP-EI. We present for the first time the cumulative regret upper bounds of GP-EI with BPMI and BSPMI. Importantly, we show that in both cases, GP-EI is a no-regret algorithm for both squared exponential (SE) and Matérn kernels. Further, we present for the first time that GP-EI with BOI either achieves a sublinear cumulative regret upper bound or has a fast converging noisy simple regret bound for SE and Matérn kernels. Our results provide theoretical guidance to the choice of incumbent when practitioners apply GP-EI in the noisy setting. Numerical experiments are conducted to validate our findings.
title Bayesian Optimization with Expected Improvement: No Regret and the Choice of Incumbent
topic Machine Learning
url https://arxiv.org/abs/2508.15674